feat(ml): introduce support of onnxruntime-rocm for AMD GPU

[Zelnes]

I'd like to propose this feature, which introduces support for machine learning on AMD GPU.

:warning: Not stable

It's relying on an opened PR which disable some caching features, in order to be able to run in parallel. (IMHO parallelizing without cache is still faster than caching in single threaded mode).

Important note

I just tried to make something work for me, and I'm not pretending to propose something completely working for anyone anywhere. I'm proposing this here, so advanced users/developers, can provide help, add some tests, and to make this available for others.

I hope I'll have some feedback :+1:

Notes

Docker size

Second note: the downside of all this new docker AMD capable, is the 28GB size of final image. I hope someone can help reduce this size.

Links

Please, see this discussion where I exchanged with @mertalev on this, and I posted more explanations, which led me here.